Responsible AI

When to use which LLM

Choosing the right tool for the task is the single biggest lever on AI energy use. Reserve heavy reasoning models for high-value work, default to efficient tiers for everyday queries, and skip AI entirely where simpler tools do the job.

Use AI

Drafting emails & meeting notes

High productivity gain, tiny model is sufficient. Tier-A models only.

Use AI

Summarising documents

Saves hours of human attention; energy per summary is negligible vs. the time saved.

Use AI

Code autocomplete & refactors

Use a small model for completion; reserve large models for architecture questions.

Use AI

Customer support triage

Route first-line responses through a Tier-A model with human escalation.

Skip

Generating images for every blog post

Image generation is ~100× more energy than text. Use only when the image earns its place.

Skip

Real-time AI video for internal demos

Highest energy class. Record once or use static assets instead.

Skip

Running 'deep research' on routine questions

Reasoning loops can use 10–30× a standard query. Save for genuinely complex problems.

Skip

AI search instead of a known URL

If you know the source, go direct. AI search is convenient but wasteful as a habit.

Careful

Multi-agent workflows

Powerful, but each step compounds energy. Cap step count and prefer smaller agents.

Careful

Fine-tuning on your data

Training is energy-intensive up-front but pays back if it lets you use a smaller model.

Use AI

Voice transcription & translation

Mature, efficient, high human value.

Careful

AI-generated slide decks end-to-end

Worth it for first drafts; don't regenerate 20 versions.

Rule of thumb

  • Tier A (Gemini Flash, Claude Haiku, GPT-4o mini): default for 70%+ of prompts — summaries, drafts, classification, extraction.
  • Tier B–C (Gemini Pro, Claude Sonnet, GPT-4o): reach for these on nuanced writing, code review, or multi-step reasoning.
  • Tier D (Claude Opus, GPT-5 / o-series): reserve for genuinely hard reasoning — research, complex code, strategy. A single deep-reasoning run can use 10–30× a standard query.